Re: [PATCH 0/1] uprobes: Kill __replace_page(), change uprobe_write_opcode() to rely on gup(WRITE)

[Oleg Nesterov <oleg@redhat.com>]

On 12/10, Linus Torvalds wrote:
>
> On Tue, Dec 10, 2013 at 11:18 AM, Oleg Nesterov <oleg@redhat.com> wrote:
> >
> > Well, if we use get_user_pages() we have vma for free.
>
> No we don't. get_user_pages() is expensive as hell.

I understand, I said "if we use get_user_pages()".

> We'd be *much* better off using get_user_pages_fast() if possible -
> and I bet _is_ possible in 99% of all cases.

We can't. get_user_pages_fast() takes mmap_sem.

Perhaps we can do something, say add FOLL_FAST_ONLY, but until then
we need the 1st get_user_pages(write => 0) anyway.

> > But I am not sure about "and verify it for now" above. Do you mean
> > that uprobe_write_opcode() should do another valid_vma() ?
>
> No. I think we should just do something like
>
>     if (vma->vm_flags & (VM_SHARED | VM_HUGETLB))
>         return -EINVAL;

OK, but WARN_ON(VM_SHARED | VM_HUGETLB) looks better imho. Because
this should not happen.

> then just looking up the page table _once_ - instead of
> three times - would make the code safer and faster.

Unlikely we can do this only once... At least I do not see how.

> It *may* be safe if callers do
> all the right black magic, but it's by no means obvious.

I agree, this needs more documentation in any case.

Oleg.